Refactor evaluation_test.py to ensure singleton watcher is initialized

- Moved the call to `ensure_singleton_watcher()` into the `evaluation_test` function to ensure the evaluation watcher is running before processing begins. This change enhances the reliability of the evaluation process by ensuring the watcher is active during execution.

Author: dphuang2

eval_protocol/pytest/evaluation_test.py

@@ -30,9 +30,6 @@
 
 from ..common_utils import load_jsonl
 
-# Ensure the evaluation watcher is running (OS-level singleton)
-ensure_singleton_watcher()
-
 
 def evaluation_test(
     *,
@@ -200,6 +197,9 @@ def wrapper_body(**kwargs):
                 eval_metadata = None
                 all_results: List[EvaluationRow] = []
 
+                # Ensure the evaluation watcher is running (OS-level singleton)
+                ensure_singleton_watcher()
+
                 try:
                     # Handle dataset loading
                     data: List[EvaluationRow] = []